AI Is Eating the Tasks That Used to Train People, and the Payroll Data Already Shows It

Artificial intelligence is entering knowledge work at the smallest unit of production, which is the task. It drafts the memo. It summarises the transcript. It generates the code stub, assembles the first deck, finds the precedent.

Job titles are unchanged. The economic content underneath them is not.

Dr Christophe Kolb, founder and chief executive of Taller and co-author of a book on working with agentic AI, argues that this creates a cost most firms have not priced. Cheaper routine output lets leaders move faster with smaller teams, at least for a while. But many of the tasks now being compressed were the training ground for professional judgment.

The first memo, the first model, the first document review, the first reconciliation. Each of those did two jobs at once. They shipped, and they taught.

The hours that survived

Research Kolb co-authored modelled five professions task by task. For every 100 hours of pre-AI work, between 57 and 73 hours survived.

The pattern matters more than the range. What remains clusters around client context, source grounding, integration, sign-off risk, stakeholder management and judgment under pressure. What went was search, drafting, summarisation, classification, coding scaffolds and first-pass analysis, all of which share three properties: digital inputs, a recognisable output format, and quality that can be reviewed.

Human value migrates accordingly, towards question selection, exception handling, persuasion, coordination and accountability. Kolb puts the consequence in one line. When answers become abundant, the scarce executive skill becomes judging which answer to trust.

They shipped, and they taught

The employment data agrees

This is no longer a projection. Stanford researchers working from ADP payroll records found that employment for 22 to 25 year olds in the most AI-exposed occupations now sits 19 per cent below where it would have been had it tracked their less-exposed peers.

Employment for experienced workers in the same occupations held steady.

That is the shape of the problem stated precisely. The technology is not reducing headcount across the board. It is removing the bottom rung.

What incidental learning actually was

The old apprenticeship model worked by accident. Juniors pulled comparables, cleaned data, built slides, checked citations, read cases, updated tickets and reconciled accounts. Most of it was tedious, and all of it was teaching pattern recognition nobody was explicitly delivering.

Kolb's examples are the useful part, because each one names a specific competence acquired through drudgery. A junior banker learned why a comparable felt wrong. A paralegal learned why a citation that looked clean failed. A developer learned how one small dependency broke a living system. An auditor learned how a variance revealed a story about a process.

None of that was on a syllabus. It came from doing the boring version enough times to develop an instinct, and the boring version is exactly what has been automated.

Designing the apprenticeship on purpose

Kolb's answer is not to preserve the drudgery artificially. It is to define the learning deliberately, since it can no longer be relied on to happen by itself.

Every AI-enabled workflow, he argues, should specify four things: question rights, review standards, escalation triggers and named ownership. Those protocols determine whether AI becomes a place where juniors are taught or one where they are bypassed.

The practical shift is in what junior time is spent on. Less producing first drafts from nothing, more comparing AI output against source material, identifying failure modes, explaining verification logic and presenting tradeoffs to senior reviewers.

He proposes a replacement ladder with four rungs. Assisted production, using AI to create drafts, code, models or summaries. Verification, checking grounding, provenance, assumptions, edge cases and compliance. Exception handling, diagnosing why a plausible answer fails. And accountable recommendation, explaining a decision under uncertainty to a client, partner or manager.

The mechanisms he suggests are concrete enough to implement this quarter. Libraries of verified work. Supervised audits in which juniors hunt for omissions and hallucinated confidence. Rotations onto exception teams. Reviews of errors, near misses and judgment calls. Simulations for the rare high-stakes cases nobody encounters often enough to learn from.

The smallest of them may be the most useful: asking juniors to write short confidence notes explaining why a given output should be trusted, revised or rejected.

That is the skill the first memo used to teach, extracted and taught on purpose, because the first memo is no longer being written by a person.